Pods as Workers, Not Agents: Rethinking the Deployment Unit for AI Agents on Kubernetes
Running AI agents on Kubernetes raises a key question: should each agent get its own Pod? The kagent project argues no—agents are bursty, short-lived, can spawn subagents, and may wait for human approval, making one Pod per agent wasteful. Agent-substrate adds a control plane to schedule logical “Actors” onto long-lived worker Pods. By Mark Silvester
Kubernetes as the Foundation for AI Agent Management Framework
In a recent CNCF blog post, Lin Sun explores the idea of rethinking the deployment unit for AI agents on Kubernetes, proposing a shift from considering Pods as the primary unit to a more nuanced approach. As agent counts increase, challenges such as isolation, identity management, access control, and observability become more pressing, necessitating a reevaluation of how these agents are deployed and managed.
Sun argues that while Pods provide isolation, a dedicated Pod for each potential agent is inefficient, especially for short-lived, bursty workloads. Additionally, agents may spawn subagents, require human interaction, or remain idle, making the traditional Pod model less suitable. To address these issues, Sun suggests introducing a control plane above Kubernetes, such as Agent Substrate, which Google introduced alongside the Agent Sandbox project.
Agent Substrate provides an isolated execution environment while managing the lifecycle and placement of AI agents onto execution workers. Kubernetes itself continues to manage Pods, Services, networking, storage, and compute, with Agent Substrate handling the logical agents' scheduling and resource management. This approach allows a fixed pool of long-lived Pods to support a significantly larger number of logical agents than would be practical with a dedicated, continuously running Pod for each agent.
By treating Pods as execution workers rather than the primary deployment model for agents, Sun's proposal offers several benefits. Identity, access control, network policies, and runtime permissions can be expressed at the agent template level, with per-agent overrides. Ownership, quotas, and billing become more manageable, and observability can be associated with the logical agent rather than the Pod.
It is important to note that this proposal does not displace Kubernetes, which remains the industry standard platform for microservices and inference workloads at scale. The open question, then, is whether the Pod should continue to be the unit of deployment, identity, and lifecycle for AI agents. Sun's exploration of this concept through kagent and Agent Substrate offers a promising path forward for managing AI agents in Kubernetes environments.
Written by urgent.news from InfoQ's reporting — not their text. Machine-written — may contain errors; check the original before relying on it.
